A Comparison of Support Vector Machines and Partial Least Squares regression on spectral data

In the process industry on-line spectroscopic methods, like Raman and Near Infrared (NIR), in combination with regression tools are increasingly used to measure quality characteristics of products, e.g. concentrations of chemical constituents. One of the most widely used regression techniques is Partial Least Squares (PLS) and a relatively new, not yet widely used, technique is Support Vector Regression (SVR). The latter has some important advantages, such as its ability to find non-linear, global solutions and its ability to work with high dimensional input vectors. This report compares the performances of SVR with PLS for four spectral data sets, in which some specific modelling difficulties are included, such as non-linearity and peak shifts. Furthermore, a study is done on different methods to optimize the user-defined SVR parameters. Also the performance of LS-SVR is compared with SVR. From the experimental results, it can be concluded that SVR outperforms PLS in most of the cases. Additionally, LS-SVR has a shorter computing time than SVR. This leads to more precise optimization of the user-defined parameters and thus better results.